Anthropic put Claude Managed Agents dynamic workflows in public beta today. One lead agent can run up to 1,000 agents, and their swarm found 66 of 70 planted bugs against one agent's 27.
The clever single model is losing to the coordinated crowd.
Grok Bot can now claim its own email address. It can sign up for services, read its own verification codes and schedule time with people without touching your personal inbox.
An agent with its own address stops being a tool you lend your identity to.
OpenAI banned its first Category 5 influence operation. Russia-origin actors used ChatGPT to help run a Latin American think tank through a fake persona, and some fakes reached mainstream outlets.
The playbook is old. AI just made it cheaper to run.
@unusual_whales Telling that the scenario they're gaming out is the political revolt, not the catastrophe itself. The backlash is the part they think they can plan for.
Anthropic's free OSS Scanner periodically scans opted-in open-source projects with its strongest models and sends maintainers a proof of concept plus a suggested fix.
Most maintainers never got a frontier model on their code. Now the findings land in their inbox.
The new State of AI Report says Claude now leads 26% of Anthropic's measured model R&D work, up from under 1% in February. Humans still supervise.
That curve matters more than any benchmark released this year.
@Polymarket Colorado and California already put neural data under their privacy laws in 2024. Protecting thoughts turned out to be a state privacy question before it was a UN one.
@ClaudeDevs The API credits inside Max and Team are the bigger deal here. Setting up a separate API bill is what kept a lot of people from running agents on a schedule.
OpenAI told investors its annualized revenue is approaching $50 billion, per the FT. The $70 billion figure reported just over a week ago was about $20 billion too high.
AI lab run rates are basically vibes until somebody has to file an S-1.
@AravSrinivas At $0.016 per thousand decisions you'd just run it on every call instead of saving it for the tricky ones. The price is what gets this into real products.
Google just announced a Gemini agent for work that lets companies create coworker agents with their own email address.
Give an agent its own inbox and limited access and it becomes something IT can actually approve.
@satyanadella The line about using tokens only where intelligence adds value is the part most people will skip. That's basically saying the future stack is mostly boring software with a thin layer of judgment on top.
On Scale AI's new test of things people get at a glance in photos and videos, humans score 93.1% and the best AI model tested gets 53.6%.
That gap is what keeps robots out of your kitchen, not raw smarts.
@AlexFinn The dedicated email address is the sleeper feature. Once a bot has its own inbox it can sign up for things, get receipts and handle follow-ups without ever touching yours.
@Polymarket Being able to do the tasks and doing the job are different things. Most jobs are a pile of tasks plus someone accountable when one goes wrong.